RESEARCH OF SPEECH RECOGNITION SYSTEMS AND METHODS

Use of the speech interface for computer devices and systems control at distance is one of possible methods of physical and psychological distance abbreviation between users and computer equipments, based on the user's voice signals. The research of automated speech recognition systems, methods and algorithms using in input speech message recognition is carried out. The classification of automated speech recognition systems is formulated in the article. This classification takes into account such features as: the dictionary size, the dependence on the speaker, the speech type, systems designation, recognition algorithm etc. The results of modern speech signals processing methods analytic review are given, namely: linear prediction of speech signal, neural networks, latent Markov models and the method of dynamic time wrap. The main objective of the linear prediction method is to define a set of prediction coefficients which provide a recognition error minimization. When using neural networks for recognition of a speech signal it is necessary to build an appropriate neural network and select of the synapses weight coefficients for minimization of errors. The method of dynamic time transformation is a method of elastic comparing of the observation vector with the saved template. The advantage of hidden Markov processes is the ability to process sequences and signals of different lengths that is a difficult task for operation with neural networks. The given review allows to estimate possibilities of the existing speech signals processing methods and to determine prospect of their mathematical apparatus application in speech signals processing tasks in systems of automated speech recognition. The article has survey character.